diff options
-rw-r--r-- | arm_compute/core/NEON/wrapper/traits.h | 42 | ||||
-rw-r--r-- | src/core/NEON/kernels/NEBitwiseAndKernel.cpp | 6 |
2 files changed, 44 insertions, 4 deletions
diff --git a/arm_compute/core/NEON/wrapper/traits.h b/arm_compute/core/NEON/wrapper/traits.h index 5cd6086c0c..0dbd90ddf8 100644 --- a/arm_compute/core/NEON/wrapper/traits.h +++ b/arm_compute/core/NEON/wrapper/traits.h @@ -40,7 +40,7 @@ struct vector_64_tag {}; /** 128-bit vector tag */ struct vector_128_tag {}; -/** Create the appropriate NEON vector given its type and size */ +/** Create the appropriate NEON vector given its type and size in terms of elements */ template <typename T, int S> struct neon_vector; // Specializations #ifndef DOXYGEN_SKIP_THIS @@ -72,6 +72,46 @@ template <> struct neon_vector<float16_t, 8>{ using type = float16x8_t; using ta template <typename T, int S> using neon_vector_t = typename neon_vector<T, S>::type; /** Helper type template to get the tag type of a neon vector */ template <typename T, int S> using neon_vector_tag_t = typename neon_vector<T, S>::tag_type; + +/** Vector bit-width enum class */ +enum class BitWidth +{ + W64, /**< 64-bit width */ + W128, /**< 128-bit width */ +}; + +/** Create the appropriate NEON vector given its type and size in terms of bits */ +template <typename T, BitWidth BW> struct neon_bitvector; +// Specializations +#ifndef DOXYGEN_SKIP_THIS +template <> struct neon_bitvector<uint8_t, BitWidth::W64>{ using type = uint8x8_t; using tag_type = vector_64_tag; }; +template <> struct neon_bitvector<int8_t, BitWidth::W64>{ using type = int8x8_t; using tag_type = vector_64_tag; }; +template <> struct neon_bitvector<uint8_t, BitWidth::W128>{ using type = uint8x16_t; using tag_type = vector_128_tag; }; +template <> struct neon_bitvector<int8_t, BitWidth::W128>{ using type = int8x16_t; using tag_type = vector_128_tag; }; +template <> struct neon_bitvector<uint16_t, BitWidth::W64>{ using type = uint16x4_t; using tag_type = vector_64_tag; }; +template <> struct neon_bitvector<int16_t, BitWidth::W64>{ using type = int16x4_t; using tag_type = vector_64_tag; }; +template <> struct neon_bitvector<uint16_t, BitWidth::W128>{ using type = uint16x8_t; using tag_type = vector_128_tag; }; +template <> struct neon_bitvector<int16_t, BitWidth::W128>{ using type = int16x8_t; using tag_type = vector_128_tag; }; +template <> struct neon_bitvector<uint32_t, BitWidth::W64>{ using type = uint32x2_t; using tag_type = vector_64_tag; }; +template <> struct neon_bitvector<int32_t, BitWidth::W64>{ using type = int32x2_t; using tag_type = vector_64_tag; }; +template <> struct neon_bitvector<uint32_t, BitWidth::W128>{ using type = uint32x4_t; using tag_type = vector_128_tag; }; +template <> struct neon_bitvector<int32_t, BitWidth::W128>{ using type = int32x4_t; using tag_type = vector_128_tag; }; +template <> struct neon_bitvector<uint64_t, BitWidth::W64>{ using type = uint64x1_t; using tag_type = vector_64_tag; }; +template <> struct neon_bitvector<int64_t, BitWidth::W64>{ using type = int64x1_t; using tag_type = vector_64_tag; }; +template <> struct neon_bitvector<uint64_t, BitWidth::W128>{ using type = uint64x2_t; using tag_type = vector_128_tag; }; +template <> struct neon_bitvector<int64_t, BitWidth::W128>{ using type = int64x2_t; using tag_type = vector_128_tag; }; +template <> struct neon_bitvector<float_t, BitWidth::W64>{ using type = float32x2_t; using tag_type = vector_64_tag; }; +template <> struct neon_bitvector<float_t, BitWidth::W128>{ using type = float32x4_t; using tag_type = vector_128_tag; }; +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +template <> struct neon_bitvector<float16_t, BitWidth::W64>{ using type = float16x4_t; using tag_type = vector_64_tag; }; +template <> struct neon_bitvector<float16_t, BitWidth::W128>{ using type = float16x8_t; using tag_type = vector_128_tag; }; +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +#endif /* DOXYGEN_SKIP_THIS */ + +/** Helper type template to get the type of a neon vector */ +template <typename T, BitWidth BW> using neon_bitvector_t = typename neon_bitvector<T, BW>::type; +/** Helper type template to get the tag type of a neon vector */ +template <typename T, BitWidth BW> using neon_bitvector_tag_t = typename neon_bitvector<T, BW>::tag_type; // clang-format on // *INDENT-ON* } // namespace traits diff --git a/src/core/NEON/kernels/NEBitwiseAndKernel.cpp b/src/core/NEON/kernels/NEBitwiseAndKernel.cpp index c1e3e1f0bc..ed83286acf 100644 --- a/src/core/NEON/kernels/NEBitwiseAndKernel.cpp +++ b/src/core/NEON/kernels/NEBitwiseAndKernel.cpp @@ -42,10 +42,10 @@ class Coordinates; namespace { -template <typename T, int S> +template <typename T> inline void bitwise_and(const T *__restrict input1, const T *__restrict input2, T *__restrict output) { - using type = typename wrapper::traits::neon_vector<T, S>::type; + using type = typename wrapper::traits::neon_bitvector<T, wrapper::traits::BitWidth::W128>::type; const type val1 = vloadq(static_cast<const T *>(input1)); const type val2 = vloadq(static_cast<const T *>(input2)); @@ -108,7 +108,7 @@ void NEBitwiseAndKernel::run(const Window &window, const ThreadInfo &info) execute_window_loop(window, [&](const Coordinates & id) { - bitwise_and<uint8_t, 16>(input1.ptr(), input2.ptr(), output.ptr()); + bitwise_and<uint8_t>(input1.ptr(), input2.ptr(), output.ptr()); }, input1, input2, output); } |